A rapid fault diagnosis method and system for a subway air compressor

Abstract

The application discloses a kind of subway air compressor quick fault diagnosis method and system, and the application is according to the running state of compressor, motor current change and reservoir pressure change to identify stop, start load, steady-state gas supply and unloading phase, use each phase normal baseline to construct phase condition residual, and with original multivariate time series data form enhanced window. Edge end convolution model outputs first fault probability and uncertainty;Only in suspected fault, high uncertainty, residual anomaly or phase switching triggers background block time series model review. Fusion module dynamically weights according to model confidence, data quality and the consistency of model and residual evidence, and adopts alarm delay, conflict reservation and timeout degradation to generate final results. The application can reduce invalid review and communication burden, improve the stability and traceability of diagnosis under variable conditions.

Description

Technical Field

[0001] This invention belongs to the field of rapid fault diagnosis technology for subway air compressors, and more specifically, relates to a rapid fault diagnosis method for subway air compressors. This invention also relates to a rapid fault diagnosis system for subway air compressors. Background Technology

[0002] The air compressor in a subway car is the source of compressed air for the braking system, door system, and other pneumatic components. Operating under high load, variable conditions, and strong vibration for extended periods, the compressor's pressure, current, temperature, and valve status can change due to component wear, pipeline leaks, overheating, or electrical malfunctions. If these faults are not identified promptly, they may result in insufficient air supply pressure, reduced braking performance, or operational disruptions.

[0003] Existing diagnostic methods for subway air compressors typically employ fixed thresholds, a single machine learning model, or a concatenated convolutional and recurrent network into an end-to-end classification model. These methods often process all operating phases with a uniform statistical distribution, failing to differentiate between different phases such as startup loading, steady-state air supply, and unloading. Normal phase transitions can cause abrupt changes in pressure, current, and temperature, which can easily be misinterpreted by the model as faults.

[0004] On the other hand, continuously uploading all windows to the complex backend model for verification would consume vehicle communication bandwidth and backend computing power; relying solely on lightweight edge models would make it difficult to identify weak anomalies across cycles. Existing technologies also lack a diagnostic mechanism that selects verification windows based on edge prediction uncertainty, phase condition residuals, and phase switching events, and dynamically adjusts fusion weights under conditions of model conflicts, communication delays, and changes in data quality. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of existing technologies, achieve rapid fault diagnosis of subway air compressors with low omission rate under varying operating conditions, significantly reduce the burden of edge-cloud communication and backend computing power consumption, and have the ability to trace diagnostic results and continuously evolve the model.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A rapid fault diagnosis method for subway air compressors includes the following steps: S1: Acquire continuous physical quantities, discrete state quantities, and data quality information of the subway air compressor; identify the current operating phase based on the compressor operating status, motor current changes, and reservoir pressure changes; calculate the phase condition residual based on the normal baseline of the corresponding operating phase; and form an enhanced feature sequence containing standardized features, phase condition residual, phase encoding, and data quality mask. S2: Construct a fixed-length sliding window and a phase-aligned window with the working condition phase switching time as the anchor point along the time axis respectively. Input the two types of windows into the first diagnostic model deployed at the edge to obtain the first fault probability, prediction uncertainty and residual abnormality. S3: Determine the background review trigger status based on the first fault probability, prediction uncertainty, residual abnormality and operating condition phase switching event; only when the review trigger condition or periodic sampling condition is met, send the review data packet with device identifier, window identifier, timestamp, enhanced feature sequence and data quality information to the background terminal, and output the second fault probability by the second diagnostic model; S4: Determine the dynamic fusion weights based on the confidence of the two diagnostic models, the data quality of the data received by each model, and the consistency between the model output and residual abnormal evidence, and generate the fusion failure probability; combine the air compressor operating status constraints and alarm hysteresis conditions to output confirmed faults, temporary warnings, or normal results, and retain edge warning records when model results conflict or background verification times out; S5: Associate storage condition phase, window identifier, first failure probability, second failure probability, dynamic fusion weight, fusion failure probability, conflict state, and inference time, and use the window confirmed by the maintenance results as a candidate update sample for data drift detection and model version update.

[0007] Preferably, in S1, the continuous physical quantities include at least several of the following: compressor pressure, pneumatic panel pressure, cyclone separator outlet pressure, drying tower pressure drop, reservoir pressure, oil temperature, and motor current; the operating condition phase includes at least the shutdown phase, start-up loading phase, steady-state gas supply phase, and unloading phase; the switching of the operating condition phase must be confirmed after the corresponding state conditions continuously meet the preset number of sampling points.

[0008] Preferably, in step S1, for each working condition phase and each continuous physical quantity, the median, lower quantile, and upper quantile are calculated based on the samples marked as normal during the training period; the absolute value of the difference between the current standardized feature and the corresponding phase median is used as the numerator, and the difference between the upper quantile and the lower quantile plus a preset positive number is used as the denominator to obtain the phase condition residual, and the residual abnormality is generated based on the weighted sum of the residuals of each dimension.

[0009] Preferably, in S2, the fixed-length sliding window contains 120 to 480 continuous sampling points; the phase alignment window is constructed from the moment when the current working condition phase was last confirmed to switch, and the part that is less than the preset length is padded and a missing mask is set, and the part that exceeds the preset length is resampled or pooled. The two types of windows are spliced ​​along the feature dimension or respectively input into the two input branches of the first diagnostic model.

[0010] Preferably, in step S2, the prediction uncertainty is calculated using the normalized binary entropy corresponding to the first fault probability; when the first fault probability reaches the warning threshold, the prediction uncertainty reaches the uncertainty threshold, the residual uncertainty reaches the residual threshold, or a phase switch of the operating condition is detected, a background review is triggered, and a window is sampled from the window that has not been triggered for background inspection according to a preset period.

[0011] Preferably, in S3, the second diagnostic model divides the enhanced feature sequence into multiple temporal blocks with overlapping regions, inputs each temporal block into a multi-layer Transformer encoder after linear projection and position encoding, and outputs the second fault probability after pooling and fully connected processing; the backend aligns the second fault probability with the corresponding edge end result according to the window identifier and timestamp.

[0012] Preferably, in step S4, the dynamic fusion weight of the first diagnostic model is obtained by multiplying the model's base weight, model confidence, data quality coefficient, and consistency coefficient between the model output and residual abnormality; the base weight of the second diagnostic model is higher than that of the first diagnostic model, and the dynamic fusion weight of the second diagnostic model is reduced when the review data packet is missing, the interpolation ratio is too high, or the time alignment error is present.

[0013] Preferably, in step S4, the fault trigger threshold is higher than the fault resolution threshold; the fault is confirmed after the fusion fault probability reaches the fault trigger threshold for a preset number of consecutive times, and the fault is resolved after it falls below the fault resolution threshold for a preset number of consecutive times; when the background review does not return within the maximum waiting time, a temporary warning is output, and the confirmation status is updated according to the window identifier after the background result arrives, but the original warning record is not deleted.

[0014] Preferably, in step S5, data drift detection is performed on the feature distribution difference between the recently confirmed normal window and the training baseline; when the drift reaches a threshold or a preset number of maintenance confirmation samples are accumulated, the second diagnostic model is updated, and the updated second diagnostic model is used as the teacher model to perform knowledge distillation on the first diagnostic model; version replacement is performed only when the candidate first diagnostic model simultaneously meets the constraints of fault recall rate, model capacity and inference latency.

[0015] On the other hand, the present invention proposes a rapid fault diagnosis system for subway air compressors, including: a data access module, a working condition phase recognition module, a phase residual construction module, a window construction module, an edge diagnosis module, a review trigger module, a background review module, a dynamic fusion module, an alarm state machine module, a model management module, and a historical record module. The data access module, working condition phase recognition module, phase residual construction module, window construction module, edge diagnosis module, review trigger module, background review module, dynamic fusion module, alarm state machine module, model management module, and historical record module are configured to perform the steps of the above method. The edge diagnostic module and the review trigger module are set on the edge computing device at the vehicle end, vehicle gateway or maintenance site, while the background review module and the model management module are set on the operation and maintenance server or cloud computing node. The edge end only uploads review data packets that meet the review trigger conditions or periodic sampling conditions, and sets a non-repeatable window identifier for each review data packet.

[0016] Compared with the prior art, the technical effects and advantages of the present invention are as follows: 1. Enhanced adaptability to changing operating conditions: Normal baselines are established for shutdown, startup loading, steady-state gas supply and unloading phases, respectively, to distinguish pressure and current deviations caused by normal operating condition switching from continuous abnormalities, thereby reducing false alarms caused by uniform thresholds and uniform statistical distributions.

[0017] 2. Communication and computing power are used on demand: The edge terminal selects the verification window based on the failure probability, probability entropy uncertainty, phase residual abnormality and phase switching event, so that high confidence normal windows do not need to be uploaded window by window, while periodic sampling checks control the risk of long-term missed detections.

[0018] 3. Stable and sustainable diagnostic status: The fusion weights change dynamically with model confidence, data quality and residual consistency, and are combined with trigger / release dual thresholds, continuous window counting, conflict retention, timeout degradation, drift detection and constrained model updates to form a traceable closed loop. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the overall fault diagnosis process of the present invention. Figure 2 This is a flowchart illustrating the dual-model training data construction and dual-model training process of the present invention. Figure 3 This is a structural diagram of the three-layer one-dimensional convolutional fault diagnosis model at the edge of the present invention; Figure 4 This is a structural diagram of the background block-based timing fault verification model of the present invention; Figure 5 This is a flowchart of the dual-model online diagnosis, result fusion, and fault status determination process of the present invention. Figure 6 This is a structural diagram of the functional modules of the rapid fault diagnosis system for subway air compressors of the present invention; Figure 7 This is a graph showing the change in training loss of the three-layer one-dimensional convolutional model of this invention. Figure 8This is a graph showing the change in training loss of the block-based time series model of the present invention; Figure 9 A bar chart comparing the performance metrics of the two diagnostic models of this invention with those of support vector machines and long short-term memory networks; Figure 10 This is a comparison of the confusion matrix of the test results of the three-layer one-dimensional convolutional model and the block-based temporal model of this invention; Figure 11 This is a flowchart illustrating the working condition phase identification, phase condition residual, and enhancement window construction process of the present invention. Figure 12 This is a closed-loop diagram of event-triggered edge-cloud verification, dynamic fusion, and model update in this invention. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without inventive effort are within the scope of protection of this invention.

[0021] Please see Figure 1-12 This invention provides a rapid fault diagnosis method for subway air compressors, comprising the following steps: S1: Acquire continuous physical quantities, discrete state quantities, and data quality information of the subway air compressor; identify the current operating phase based on the compressor operating status, motor current changes, and reservoir pressure changes; calculate the phase condition residual based on the normal baseline of the corresponding operating phase; and form an enhanced feature sequence containing standardized features, phase condition residual, phase encoding, and data quality mask. S2: Construct a fixed-length sliding window and a phase-aligned window with the working condition phase switching time as the anchor point along the time axis respectively. Input the two types of windows into the first diagnostic model deployed at the edge to obtain the first fault probability, prediction uncertainty and residual abnormality. S3: Determine the background review trigger status based on the first fault probability, prediction uncertainty, residual abnormality and operating condition phase switching event; only when the review trigger condition or periodic sampling condition is met, send the review data packet with equipment identifier, window identifier, timestamp, enhanced feature sequence and data quality information to the background end, and the second fault probability is output by the second diagnostic model; S4: Determine the dynamic fusion weights based on the confidence of the two diagnostic models, the data quality of the data received by each model, and the consistency between the model output and residual abnormal evidence, and generate the fusion failure probability; combine the air compressor operating status constraints and alarm hysteresis conditions to output confirmed faults, temporary warnings, or normal results, and retain edge warning records when model results conflict or background verification times out; S5: Associate storage condition phase, window identifier, first failure probability, second failure probability, dynamic fusion weight, fusion failure probability, conflict state, and inference time, and use the window confirmed by the maintenance results as a candidate update sample for data drift detection and model version update.

[0022] like Figure 1 , Figure 5 , Figure 11 and Figure 12 As shown in this embodiment: S1: Acquire multivariate time series data and data quality information, identify the current operating condition phase, calculate the phase condition residual, and form an enhanced feature sequence; S2: Construct a fixed-length window and a phase-aligned window, and output the first fault probability, prediction uncertainty and immediate preliminary judgment result from the edge-end CNN3 model; S3: Perform a verification gating based on the fault probability, uncertainty, phase residual abnormality and phase switching event, send the trigger window to the background PatchTST model, and output the second fault probability; S4: Dynamically calculate the fusion weights based on model confidence, data quality, and consistency between the model and residual evidence, and generate the final diagnostic results by combining operational constraints, alarm hysteresis, conflict retention, and timeout degradation. S5: Save diagnostic links and window evidence, and use maintenance confirmation results for drift detection, background model updates, and edge model constrained knowledge distillation.

[0023] 1. Overall Diagnostic Methods like Figure 1 , Figure 5 , Figure 11 and Figure 12 As shown, this embodiment provides a method for diagnosing faults in subway air compressors based on operating condition phase perception and event-triggered edge-cloud verification, including the following steps: S1: Acquire multivariate time series data and data quality information, identify the current operating condition phase, calculate the phase condition residual, and form an enhanced feature sequence; S2: Construct a fixed-length window and a phase-aligned window, and output the first fault probability, prediction uncertainty and immediate preliminary judgment result from the edge-end CNN3 model; S3: Perform a verification gating based on the fault probability, uncertainty, phase residual abnormality and phase switching event, send the trigger window to the background PatchTST model, and output the second fault probability; S4: Dynamically calculate the fusion weights based on model confidence, data quality, and consistency between the model and residual evidence, and generate the final diagnostic results by combining operational constraints, alarm hysteresis, conflict retention, and timeout degradation. S5: Save diagnostic links and window evidence, and use maintenance confirmation results for drift detection, background model updates, and edge model constrained knowledge distillation.

[0024] In S1, continuous physical quantities include at least several of the following: compressor pressure, pneumatic panel pressure, cyclone separator outlet pressure, drying tower pressure drop, reservoir pressure, oil temperature, and motor current; operating phases include at least the shutdown phase, startup loading phase, steady-state gas supply phase, and unloading phase; operating phase switching must be confirmed after the corresponding state conditions continuously meet the preset number of sampling points. For each operating phase and each dimension of continuous physical quantity, the median, lower quantile, and upper quantile are calculated based on the samples marked as normal during the training period; the absolute value of the difference between the current standardized feature and the corresponding phase median is used as the numerator, and the difference between the upper and lower quantiles plus a preset positive number is used as the denominator to obtain the phase condition residuals, and the residual constant is generated based on the weighted sum of the residuals of each dimension.

[0025] 2. Acquisition of multivariate time series data This embodiment collects continuous physical quantities and discrete state quantities generated during the operation of the subway air compressor. Data sources include vehicle network, maintenance equipment interface, historical operation and maintenance database, or data files to be diagnosed, and the sampling frequency is set to 1Hz.

[0026] The continuous physical quantities include at least: TP2 compressor pressure; TP3 pneumatic panel pressure; H1 cyclone separator outlet pressure; Pressure drop in the drying tower; Reservoir pressure; Engine oil temperature; Motor current.

[0027] Discrete state quantities include at least: COMP compressor operating status; LPS low-pressure protection status; Pressure switch status.

[0028] The aforementioned continuous physical quantities and discrete state quantities together constitute the eigenvector at time t. In this embodiment, the input feature dimension is d=10.

[0029] 3. Data Preprocessing 3.1 Finite Value Filtering and Standardization First, check the original time series data for null values, positive infinity values, negative infinity values, and values ​​that are obviously out of bounds. For a small number of missing data, interpolation between adjacent time points, median padding, or forward padding can be used to repair them; for data segments with consecutive missing times exceeding a preset threshold, they are marked as invalid data segments and excluded.

[0030] To eliminate the influence of different sensor dimensions and numerical ranges, statistics within the training time period are used to standardize continuous physical quantities. The standardization result of the j-th dimension feature at time t is calculated according to equation (1):

[0031] In the formula, This represents the original value of the j-th dimension feature at time t; Represents the standardized eigenvalues; The standard deviations in the denominator of equation (1) represent the mean and standard deviation of the j-th feature during the training period, respectively. This indicates a preset positive number to prevent the denominator from being zero.

[0032] The mean and standard deviation are calculated only based on data from the training period and applied to validation data, test data, and online monitoring data to avoid future information leakage.

[0033] 3.2 Construction of Effective Fault Tags To reduce invalid alarms generated when the air compressor is stopped or in standby mode, an effective fault label is constructed by combining the air compressor's operating status and discrete alarm status.

[0034] Let the integrated alarm state represent the discrete alarm state at time t. abnormal, Anomaly, LSP protection triggered or When at least one of the exceptions is triggered, make Otherwise .

[0035] The valid fault label at time t is determined according to formula (2):

[0036] In the formula, This indicates an indicator function that takes the value 1 if the condition is true, and 0 otherwise. This represents the operating status of the air compressor at time t; Indicates the overall alarm status; This indicates the final valid fault label generated.

[0037] When the air compressor is in active operation and at least one alarm condition is triggered, Otherwise, .

[0038] 3.3 Sliding Window Construction After completing data cleaning, standardization, and fault label construction, the multivariate time series data are sorted according to the data collection time, and a fixed-length overlapping sliding window is constructed along the time axis.

[0039] Using time t as the end time of the window, the corresponding time series window is constructed according to equation (3):

[0040] In the formula, Let L represent the feature vector at time t; L represents the number of consecutive sampling points contained in a sliding window. This represents a multivariate time series window ending at time t.

[0041] This embodiment is configured as follows:

[0042] In the formula, S represents the step size between adjacent sliding windows. Therefore, the size of each model input sample is 240×10.

[0043] The dataset should first be divided into training, validation and test sets according to the collection time order, and then a sliding window should be generated within each dataset to prevent duplicate sampling points in adjacent windows from appearing in different datasets at the same time.

[0044] 3.4 Operating Condition Phase Recognition like Figure 11 As shown, this embodiment does not treat all running phases as the same statistical distribution. Instead, based on the COMP compressor operating status, motor current change, and reservoir pressure change rate, the air compressor operation process is divided into shutdown phase, start-up loading phase, steady-state air supply phase, and unloading phase.

[0045] When COMP is in a non-operating state and the motor current is below the no-load threshold, it is identified as a shutdown phase; when COMP switches from a non-operating state to an operating state, or when the motor current and reservoir pressure change rate reach the start-up threshold, it is identified as a start-up loading phase; when COMP remains operating and the current and pressure changes are within the corresponding stable range, it is identified as a steady-state gas supply phase; when COMP switches from an operating state to a non-operating state, or when the motor current continues to decrease, it is identified as an unloading phase.

[0046] To avoid frequent phase switching caused by state variable jitter, phase switching is only confirmed when a new phase condition continuously meets a preset number of sampling points, and the confirmation moment is used as the anchor point of the phase alignment window. The phase conditions and thresholds for different vehicle models can be calibrated based on the control logic and normal operation data.

[0047] Phase determination can be calibrated according to vehicle model and sensor configuration: During the training period, the system is first segmented based on the switching points of the compressor's operating state. Then, using confirmed normal motor current and reservoir pressure sequences, the distribution of shutdown current, starting increase change, steady-state fluctuation range, and unloading decrease change are statistically analyzed. Candidate thresholds for each phase are formed using corresponding empirical quantiles, and the thresholds and the number of continuously confirmed sampling points are adjusted during the verification period. The vehicle model, unit, sampling frequency, and preprocessing version corresponding to the thresholds are saved. When phase criteria conflict or there are insufficient normal samples, the phase is marked as uncertain, and a global normal baseline is used for backtracking and a verification is triggered.

[0048] 3.5 Phase Condition Residual and Residual Stability For each phase under each operating condition and for each continuous physical quantity, the phase median, lower quantile, and upper quantile are calculated from the data labeled as normal within the training time period. The phase conditional residual corresponding to the feature at the current moment is calculated using the following formula:

[0049] in This represents the phase condition residual of the i-th feature at time t under the current operating condition phase s. It is the sensor feature value of the j-th dimension at time t after standardization. It is the median of the jj-th dimension of the normal samples under operating condition phase s. and These are the upper and lower quantiles of the j-th dimension normal sample under this phase, respectively. It is a preset positive number to prevent the denominator from being zero; In the formula, the numerator includes the standardized sensor characteristics and the median corresponding to the current operating condition phase; the denominator includes the difference between the upper and lower quantiles corresponding to the current operating condition phase, as well as a positive number to prevent the denominator from being zero. The phase condition residual is normalized using the normal fluctuation range within the phase, so that the current rise and pressure change that should occur during the start-up loading phase are not directly equivalent to the abnormal deviation during the steady-state gas supply phase.

[0050] Non-negative weights are assigned based on the different sensors' sensitivities to faults, and the phase conditional residuals of each dimension are mapped to residual constants between 0 and 1:

[0051] Let represent the residual constant obtained by integrating the phase condition residuals of each dimension at time t, with values ​​between 0 and 1, and exp represent an exponential function with the natural constant as the base. This represents the summation over all d-dimensional continuous physical quantities. This represents the non-negative weight assigned to the j-th continuous physical quantity based on its sensitivity to faults. Let represent the phase condition residual of the j-th dimension feature at time t under the current operating phase.

[0052] When there are insufficient normal samples for a certain phase, the adjacent phase baseline or the global normal baseline is used as the backoff baseline; quality masks are set synchronously for missing, interpolated or out-of-bounds sampling points to avoid the values ​​being filled in to obtain the same fusion weight as the true sampled values.

[0053] 3.6 Fixed Window and Phase Alignment Window A fixed-length window is used to preserve continuous local changes; the phase alignment window starts from the most recently confirmed phase switching anchor point and ends at the current time. If the phase alignment window is less than the preset length, it is padded and a mask is set; if it exceeds the preset length, piecewise averaging, interpolation resampling, or adaptive pooling are performed.

[0054] The two types of windows can be concatenated along the feature dimension, or they can be input separately into a dual-branch feature extractor and then merged. Enhanced diagnostic samples include at least standardized features, phase conditional residuals, phase encoding, and quality masks, thereby enabling the model to simultaneously obtain absolute state, relative phase shift, and data reliability.

[0055] The original model comparison experiment in the paper uses a 240×10 basic input, namely 7 continuous physical quantities and 3 discrete state quantities. When using the enhanced input of this implementation, the phase conditional residuals of the 7 continuous quantities can be calculated separately, and quality masks can be generated for the 10 original inputs. If the phase uses 4-dimensional one-hot encoding, the dimension of a single branch input is 10+7+10+4=31, and the fixed window and the phase-aligned window are concatenated according to the feature dimension, resulting in 240×62. When using a dual-branch structure, each branch is 240×31. The number of input channels in the first layer of CNN3 and the dimension of the PatchTST linear projection input must be adjusted to correspond to the selected enhanced input method; the weights of the 240×10 basic model cannot be directly used without adaptation.

[0056] In S2, a fixed-length sliding window contains 120 to 480 continuous sampling points; the phase alignment window is constructed starting from the most recently confirmed switching moment of the current operating condition phase, padding the part that is less than the preset length and setting a missing mask, and performing resampling or pooling on the part that exceeds the preset length. The two types of windows are spliced ​​along the feature dimension or respectively input into the two input branches of the first diagnostic model; the prediction uncertainty is calculated using the normalized binary entropy corresponding to the first fault probability; when the first fault probability reaches the warning threshold, the prediction uncertainty reaches the uncertainty threshold, the residual abnormality reaches the residual threshold, or a switching of the operating condition phase is detected, a background review is triggered, and windows that have not been reviewed are sampled in the background for random checks according to a preset period.

[0057] 4. CNN3 Edge Fast Diagnostic Model like Figure 3 As shown, the first diagnostic model uses a three-layer one-dimensional convolutional network CNN3. The input window shape is 240×10. Before performing the convolution operation, the input data is converted into a "channel × time" arrangement.

[0058] The first 1D convolutional layer has 64 output channels and a kernel length of 7; the second 1D convolutional layer has 128 output channels and a kernel length of 5; the third 1D convolutional layer has 128 output channels and a kernel length of 5.

[0059] Each convolutional layer employs padding to maintain the specified time. The first two convolutional layers are followed by a GELU activation function and a Dropout layer to enhance non-linear feature extraction and reduce the risk of overfitting.

[0060] The output of the third convolutional layer is processed by adaptive average pooling and a fully connected layer, and finally the first fault probability is output through the Sigmoid function. .

[0061] The CNN3 model primarily extracts the variation patterns of pressure, current, temperature, and operating status over a local time range. Due to its lightweight network structure, it can be deployed on vehicle terminals, in-vehicle gateways, or on-site maintenance computers for low-latency initial fault diagnosis.

[0062] When the first failure probability satisfies the following formula, an immediate failure warning is generated at the edge:

[0063] In the formula, This represents the first warning threshold for the CNN3 model. This refers to the first failure probability output by the first diagnostic model at the edge (CNN3 model).

[0064] To determine whether the edge model requires backend review, the prediction uncertainty is calculated using the normalized binary entropy corresponding to the first failure probability:

[0065] This represents the prediction uncertainty of the first diagnostic model at the edge (CNN3 model). ln represents the natural logarithm function. The denominator ln2 is used to normalize the binary entropy to between 0 and 1, so that the closer the prediction probability is to 0.5, the closer the uncertainty is to 1, and the closer it is to 0 or 1, the closer the uncertainty is to 0.

[0066] The review trigger state is determined by the first failure probability, prediction uncertainty, residual abnormality, and phase switching event.

[0067] This indicates the background review trigger status at time t (a value of 1 triggers the review and uploads the data packet, while a value of 0 does not trigger it). This indicates an indicator function (1 if the condition is true, 0 otherwise). Indicates the edge warning threshold. Indicates the uncertainty threshold. Indicates the residual abnormal threshold. and These represent the operating phases at time t and time t-1, respectively. V represents logical OR, meaning that any one of the four conditions (suspected fault, high uncertainty, residual abnormality, or phase switching) will trigger a background review.

[0068] In the formula, the three thresholds are, in order, the edge warning threshold, the uncertainty threshold, and the residual abnormality threshold. When the trigger function is set to 1, a review data packet is uploaded; when the trigger function is set to 0, the edge results are retained without calling the background model window by window. To prevent long-term reliance solely on the edge model, background sampling is performed on windows that have never triggered review, according to a preset cycle.

[0069] In S3, the second diagnostic model divides the enhanced feature sequence into multiple temporal blocks with overlapping regions. Each temporal block is then input into a multi-layer Transformer encoder after linear projection and position encoding. After pooling and fully connected processing, the second fault probability is output. The backend aligns the second fault probability with the corresponding edge end result based on the window identifier and timestamp. In S4, the dynamic fusion weight of the first diagnostic model is obtained by multiplying the model's base weight, model confidence, data quality coefficient, and consistency coefficient between the model output and residual abnormality. The base weight of the second diagnostic model is higher than that of the first diagnostic model. Furthermore, the dynamic fusion weight of the second diagnostic model is reduced when there are missing data packets, excessively high interpolation ratios, or time alignment errors.

[0070] 5. PatchTST high-precision verification model in the background like Figure 4 As shown, the second diagnostic model uses the PatchTST block-based temporal model. The PatchTST model receives a temporal window that is time-aligned with the CNN3 model and divides the temporal window of length L into multiple temporal blocks with overlapping regions.

[0071] In this embodiment, the block length P=16 and the block step size Q=8. The number of blocks is calculated according to the following formula:

[0072] When L=240, P=16, and Q=8, we get .

[0073] This indicates the number of time series blocks obtained by the PatchTST block timing model in the background after dividing a time series window of length L. L represents the number of consecutive sampling points contained in the input time series window, P represents the length of each time series block, and Q represents the block step size between adjacent time series blocks.

[0074] Each temporal block is linearly projected into a fixed-dimensional embedding vector and then positional encoding is superimposed. Subsequently, the embedding sequence is input into a Transformer encoder consisting of a multi-head self-attention layer, a feedforward network, residual connections, and a normalization layer.

[0075] The Transformer encoder output is processed by average pooling and a fully connected layer, and the second fault probability is output through the Sigmoid function. .

[0076] When the second failure probability satisfies the following formula, the backend confirms that the corresponding window has a failure risk:

[0077] In the formula, This represents the second diagnostic threshold of the PatchTST model. This represents the second fault probability output by the second diagnostic model (PatchTST block timing model) on the backend.

[0078] The PatchTST model can extract air compressor pressure cycles, current start-stop, temperature trends, and long-range correlations between different sensors, so it is preferred to deploy it on a back-end server, data center, or cloud computing node.

[0079] The backend is not required to continuously receive all raw windows. The verification data packet must at least carry the device identifier, unique window identifier, window start and end times, model and preprocessing version, normalized window, phase conditional residual, phase encoding, and quality mask. The backend first verifies the field order, sampling frequency, window length, and version information; if verification fails, it returns an invalid data status and does not fuse the result with the marginal probabilities. If compressed representation is used for transmission, the backend should reconstruct a complete window consistent with its training input using the same preprocessing version before performing PatchTST inference.

[0080] After completing inference, PatchTST returns the second failure probability with the same window identifier. The fusion module aligns the outputs of the two models according to the window identifier rather than the return order, thus adapting to network jitter, queuing delays, and asynchronous returns. Windows that do not return within the maximum waiting time enter timeout degradation processing.

[0081] The maximum background waiting time is determined by the historical distribution of round-trip latency, background queuing latency, and single-window inference latency of the target deployment network, and is configured along with the device identifier and version. After a timeout, a temporary alert is retained at the edge; when the background result is late, only a review record and status change are added according to the window identifier, without overwriting the original alert time and the original model output.

[0082] 6. Model Training like Figure 2 As shown, the model training process includes, in sequence, the following steps: access to raw monitoring data, data cleaning, construction of effective fault labels, generation of sliding window samples, training of CNN3 model, training of PatchTST model, testing and evaluation, selection of thresholds, and deployment of the model.

[0083] The training set, validation set, and test set are divided according to the collection time, and the proportions can be set to 60%, 20%, and 20%, respectively.

[0084] In this embodiment, when using publicly available MetroPT3 data, the training data is taken from February 1 to May 31, 2020; the validation data is taken from June 1 to June 30, 2020; and the test data is taken from July 1 to August 31, 2020. The data should be divided into time periods first, and then a sliding window should be constructed independently within each period; windows that cross the division boundaries should not be used for training or evaluation.

[0085] When the number of normal samples significantly exceeds the number of faulty samples, negative sample downsampling, class weighting, or focus loss can be used to address class imbalance. The focus loss is calculated using the following formula:

[0086] In the formula, Indicates the focal loss value. This represents the model's predicted probability of the true class. Indicates category weight; represents the focusing parameter; ln represents the natural logarithm function.

[0087] In this embodiment, the focusing parameters can be set as follows: The parameters corresponding to the basic model experiments in the paper are as follows: window length 240 sampling points, step size 30 sampling points, training epochs 12, batch size 256, initial learning rate 0.0005, weight decay 0.01, ratio of normal samples to faulty samples during training sampling 20:1, and focus loss focusing parameter 2.0. The model checkpoints and the first and second classification thresholds were selected solely based on validation data; test data was not used in parameter selection.

[0088]

[0089] The model employs an early stopping strategy on the validation set and selects model parameters based on the fault recall rate, overall score, and validation loss on the validation set, rather than directly using the model parameters corresponding to the last training round.

[0090] 7. Experimental Results and Performance Analysis To compare the performance of basic diagnostic models, under the same conditions of data preprocessing, label construction, and time-sequence partitioning, comparative experiments were conducted using SVM, LSTM, CNN3, and PatchTST models, respectively.

[0091] Using precision, recall, accuracy and... The performance of the model is evaluated by a score, and the calculation formula is as follows:

[0092] In the formula, TP represents the number of faulty samples correctly identified as faulty; TN represents the number of normal samples correctly identified as normal; FP represents the number of normal samples incorrectly identified as faulty; FN represents the number of faulty samples incorrectly identified as normal; Precision represents precision; Recall represents recall; and Accuracy represents accuracy. It represents the harmonic mean composite score of precision and recall.

[0093] 7.1 Training Results of CNN3 Model like Figure 7 As shown, the training loss of the CNN3 model generally decreases as the training process progresses, indicating that the three-layer one-dimensional convolutional network can gradually learn the local temporal features in the multivariate data of the air compressor.

[0094] The validation loss of the CNN3 model decreases significantly in the early stages of training, but fluctuates somewhat in subsequent training. To reduce the impact of overfitting, this embodiment selects the optimal model parameters based on the validation set fault recall, overall score, and validation loss.

[0095] The single-window inference latency of the CNN3 model is approximately 5.3ms. This value represents the model inference time in the RTX 3090 server environment described in the paper, and does not include acquisition, transmission, and alarm generation time.

[0096] 7.2 Training Results of the PatchTST Model like Figure 8 As shown, the training loss of the PatchTST model decreases rapidly in the early stages of training and then remains at a low level, indicating that the model is able to learn cross-period associations in multivariate long time series data.

[0097] The validation loss of the PatchTST model is higher than the training loss and fluctuates to some extent. This is due to factors such as the imbalance between the number of normal samples and faulty samples, the differences in data distribution across different runtime stages, and the complexity of faulty sample types.

[0098] Therefore, this embodiment employs class weights, focus loss, early stopping on the validation set, and diagnostic threshold selection to reduce the impact of class imbalance and model overfitting on diagnostic results.

[0099] The single-window inference latency of the PatchTST model is approximately 18.7ms. This value represents the model inference time within the same server environment and does not include data upload, queuing, and return latency. When deployed in the background, it can be used for high-precision verification of the warning windows generated by the CNN3 model.

[0100] 7.3 Comparison of Model Performance Indicators On the validation set, the classification thresholds are scanned. First, candidate thresholds are selected based on the failure recall rate reaching the preset lower limit. Then, precision and accuracy are selected from the candidate thresholds. The highest threshold for the score or weighted composite score is used to reduce the risk of missed fault detection.

[0101] Under the same evaluation metrics, the SVM, LSTM, CNN3, and PatchTST models were tested respectively, and the experimental results are shown in Table 1 and 2. Figure 9 As shown.

[0102] Table 1 and Figures 7 to 10 The table reflects the basic model experiments presented in the paper. The total number of test samples for the two models in the confusion matrix is ​​not the same: CNN3 has 10,978 normal samples and 3,783 faulty samples, while PatchTST has 26,146 normal samples and 3,790 faulty samples. Therefore, the indicators in the table are used to illustrate the performance of each model on the corresponding test data, and the two confusion matrices are not considered as the same batch of window-paired samples.

[0103] Table 1. Comparison of Performance Indicators of Different Models

[0104] From Table 1 and Figure 9 It can be seen that the precision, recall, accuracy, and overall score of the SVM model are 0.60, 0.43, 0.68, and 0.50, respectively; while the corresponding metrics for the LSTM model are 0.65, 0.78, 0.73, and 0.71, respectively.

[0105] The CNN3 model achieved a precision of 0.70, a recall of 0.96, an accuracy of 0.78, and an overall score of 0.81. The CNN3 model exhibits a high recall rate, and the inference latency for a single time window is approximately 5.3ms; this time represents the model inference time in the stated server environment, and its performance on edge devices requires further verification.

[0106] The PatchTST model achieved a precision of 0.72, a recall of 0.97, an accuracy of 0.80, and an overall score of 0.83, outperforming other comparable models across all metrics. The inference latency for a single time window of the PatchTST model is approximately 18.7ms; this time refers only to the model inference time within the stated server environment and does not include network round-trip and queuing times.

[0107] The above basic model results show that the CNN3 model and the PatchTST model have advantages in terms of real-time diagnosis and diagnostic accuracy, respectively, providing experimental basis for the dual-model collaborative diagnosis scheme of this invention, which combines rapid initial judgment at the edge end with high-precision verification at the back end.

[0108] Figures 7 to 10 The basic window model experiments corresponding to Table 1 in the paper did not measure the gains and end-to-end latency of the newly added working condition phase residuals, event triggering, dynamic fusion, and model update mechanisms in real vehicle deployment; the parameters and effects of this part of the implementation need to be verified separately according to the target vehicle model.

[0109] 7.4 Analysis of Confusion Matrix Results like Figure 10 As shown, in the test samples corresponding to the CNN3 model, 8059 normal samples were correctly identified and 2919 were incorrectly identified as faulty; 3632 faulty samples were correctly identified and 151 were incorrectly identified as normal.

[0110] The fault recall rate of the CNN3 model is:

[0111] This formula represents the fault recall rate of the CNN3 model. The numerator 3632 represents the number of faulty samples correctly identified as faulty (TP), and the denominator 3632 represents TP, while 151 represents the number of faulty samples incorrectly identified as normal (FN). This formula is derived from... Calculate the recall rate of the CNN3 model for faulty samples.

[0112] The results show that the CNN3 model misses only a small number of faulty samples, which meets the requirement of low false negative rate for edge fault early warning. Since it identifies some normal samples as faulty, the output of the CNN3 model is preferred as the immediate early warning result and is further reviewed by the background model.

[0113] In the test samples corresponding to the PatchTST model, 20,692 normal samples were correctly identified and 5,454 were incorrectly identified as faulty; 3,676 faulty samples were correctly identified and 114 were incorrectly identified as normal.

[0114] The recall rate of the PatchTST model is:

[0115] This represents the fault recall rate of the PatchTST model. The numerator 3676 represents the number of faulty samples correctly identified as faulty (TP), and the denominator 3676 represents TP, while 114 represents the number of faulty samples incorrectly identified as normal (FN). This formula is used... Calculate the recall rate of the PatchTST model for faulty samples. The results show that the PatchTST model misses a small number of fault samples and can be used as a high-precision background verification model.

[0116] In S4, the fault trigger threshold is higher than the fault resolution threshold; the fault is confirmed after the fusion fault probability reaches the fault trigger threshold for a preset number of consecutive times, and the fault is resolved after it falls below the fault resolution threshold for a preset number of consecutive times; when the background review does not return within the maximum waiting time, a temporary warning is output, and the confirmation status is updated according to the window identifier after the background result arrives, but the original warning record is not deleted.

[0117] 8. Event-triggered review, dynamic fusion, and alarm state machine During online diagnosis, the edge terminal outputs a first fault probability for each enhanced window; for windows that meet the triggering condition or periodic sampling condition, the backend terminal outputs a second fault probability. The model confidence is calculated based on the distance between the output probabilities of the two models and the uncertainty probability of 0.5.

[0118] In the formula, This represents the first confidence level of the CNN3 model. This represents the second confidence level of the PatchTST model. This represents the first failure probability output by the CNN3 model. This represents the second failure probability output by the PatchTST model. The absolute value sign is used to calculate the distance between the output probability and the uncertainty probability 0.5. The closer the model output is to 0 or 1, the higher the confidence level; the closer it is to 0.5, the lower the confidence level.

[0119] Further, the residual consistency coefficient of each diagnostic model is calculated based on the residual abnormality, and the dynamic fusion weight is calculated by combining the basic weights, model confidence, and data quality coefficient of the data received by the model:

[0120] represents the residual consistency coefficient of the i-th diagnostic model (the degree of closeness between the model failure probability and the abnormal evidence of phase residuals), and represents the failure probability output by the i-th diagnostic model. This represents the residual abnormality at time t. This represents the dynamic fusion weights of the i-th diagnostic model. This represents the basic weights of the model. This indicates the confidence level of the model. This represents the data quality coefficient of the data received by the i-th diagnostic model at time t. It is used to evaluate the reliability of the model's input data based on factors such as the effective sampling ratio, interpolation ratio, missing mask, and time alignment error. To prevent the denominator from being zero or the value from being too small, a preset positive number is used.

[0121] In the formula, the residual consistency coefficient represents the degree of closeness between the model failure probability and the evidence of phase residual anomalies; the data quality coefficient is calculated based on the effective sampling ratio, interpolation ratio, missing mask, and time alignment error. When the verification data is missing or the returned time alignment error increases, the dynamic fusion weight of the background model is reduced accordingly.

[0122] The data quality coefficient can be calculated from the proportion of original valid samples, the proportion of interpolated samples, and the proportion of time alignment errors within the corresponding input window: original valid samples are weighted fully, interpolated samples are weighted according to the reduced weight selected for the validation data, and then the proportion of time alignment errors is subtracted, with the resulting value truncated to 0 to 1. If a model's input window has no valid samples or its data packet verification fails, the model's fusion weight is reset to zero; if both models are unavailable, the output data is invalid, and no fault determination is made.

[0123]

[0124] In the formula, Indicates the probability of fusion failure. This represents the dynamic fusion weights of the CNN3 model. This represents the first failure probability output by the CNN3 model. This represents the dynamic fusion weights of the PatchTST model. This represents the second failure probability output by the PatchTST model. The formula obtains the final fused failure probability by dynamically weighting the failure probabilities of the two models.

[0125] The left side of the equals sign represents the probability of fusion failure. Under normal communication and data integrity conditions, the base weights of the PatchTST model are preferably higher than those of the CNN3 model; however, the final weights are not fixed, but change dynamically with confidence, data quality, and residual consistency.

[0126] The alarm state machine sets fault trigger thresholds and fault clearing thresholds, with the fault trigger threshold being higher than the fault clearing threshold. A fault is confirmed when the probability of a fault reaches the trigger threshold for a consecutive preset number of windows, and the fault is cleared when the probability falls below the clearing threshold for a consecutive preset number of windows. This avoids repeated alarm switching caused by probability fluctuations around a single threshold.

[0127] When the edge end and the backend end make inconsistent judgments, retain the edge end's real-time warning and original window evidence, and mark the window as conflict pending confirmation; when the backend review times out, output a temporary warning and do not delete the edge record, and update the confirmation status according to the window label after the backend result is returned.

[0128] When COMP is not in operation, the system does not directly output confirmation of air compressor operation faults, but still retains a high residual window to identify sensor drift, pressure leakage or abnormal changes after shutdown; thus, operation fault constraints and equipment abnormality traces are processed separately.

[0129] In S5, data drift detection is performed on the feature distribution differences between the recently confirmed normal window and the training baseline. When the drift reaches the threshold or a preset number of maintenance confirmation samples are obtained, the second diagnostic model is updated, and the updated second diagnostic model is used as the teacher model to perform knowledge distillation on the first diagnostic model. Version replacement is performed only when the candidate first diagnostic model simultaneously meets the constraints of fault recall rate, model capacity and inference latency.

[0130] 9. Data drift detection and constrained model update The model management module compares the feature distribution differences between the recently confirmed normal window and the training baseline at preset intervals. The distribution differences can be calculated using the overall stability index, the maximum mean difference, or the phase quantile offset, and the drift degree of each operating condition is recorded separately.

[0131] For each vehicle model and operating condition phase, a fixed training baseline is established: statistical intervals for each continuous feature are divided using the quantiles of the training data, and the proportion of normal samples in each interval is recorded; the proportion of the same interval is statistically calculated for the recent window, and the drift is calculated based on the difference between the two proportions. The drift threshold is calibrated using historical verification periods without observed faults, and the candidate update process only begins when the threshold is exceeded for multiple consecutive monitoring cycles. This drift is only used to trigger manual review or candidate training and does not directly generate fault labels.

[0132] When the drift of any phase under any operating condition reaches a threshold, or when a preset number of maintenance confirmation samples are accumulated, the backend establishes a candidate training set to update the PatchTST model. Maintenance confirmation samples include confirmation windows after fault repair, false alarm windows that have been cleared, and conflict windows that have been manually reviewed. Windows that are only output by the model itself with high confidence but have not been confirmed by rules or maintenance results are not used as the sole supervision labels.

[0133] The updated PatchTST model serves as the teacher model, jointly supervising the marginal CNN3 candidate models with soft probabilities and true labels. A new marginal model version is generated only when the candidate model simultaneously satisfies the constraints of minimum fault recall, maximum model capacity, and maximum single-window inference latency on the independent time-series validation set.

[0134] The new version is deployed using a gray-scale approach while retaining the previous version. If the periodic sampling discrepancies, conflict rates, or missed detection indicators of the new version exceed the rollback threshold, the model management module will revert to the previous version. The model version, training time range, confirmed sample source, and validation results are all written to the audit log to prevent untraceable performance degradation caused by online self-updates.

[0135] The training samples for candidate models must be associated with maintenance work orders, retest results, and the original window identifier. An independent validation period following the training period is used to compare the recall rate, false positive rate, and single-window inference latency between the old and new versions; the official version will not be replaced if any metric fails to meet the pre-defined business constraints. During canary releases, the weights, thresholds, and preprocessing configurations of the previous version are retained for rollback purposes.

[0136] On the other hand, the present invention proposes a rapid fault diagnosis system for subway air compressors, including: a data access module, a working condition phase recognition module, a phase residual construction module, a window construction module, an edge diagnosis module, a review trigger module, a background review module, a dynamic fusion module, an alarm state machine module, a model management module, and a historical record module. The data access module, working condition phase recognition module, phase residual construction module, window construction module, edge diagnosis module, review trigger module, background review module, dynamic fusion module, alarm state machine module, model management module and history record module are configured to execute the steps of the above-mentioned rapid fault diagnosis method for subway air compressors. The edge diagnostic module and the review trigger module are set on the edge computing device at the vehicle end, vehicle gateway or maintenance site, while the background review module and the model management module are set on the operation and maintenance server or cloud computing node. The edge end only uploads review data packets that meet the review trigger conditions or periodic sampling conditions, and sets a non-repeatable window identifier for each review data packet.

[0137] 10. System Implementation like Figure 6 and Figure 12 As shown, the subway air compressor fault diagnosis system includes a data access module, a working condition phase identification module, a phase residual construction module, a window construction module, an edge diagnosis module, a review trigger module, a background review module, a dynamic fusion module, an alarm state machine module, a model management module, a historical record module, and a trend and root cause prompting module.

[0138] The operating condition phase identification module identifies the current phase based on changes in COMP, motor current, and reservoir pressure; the phase residual construction module calls the sub-phase normal baseline library to generate residual abnormality; the review triggering module only constructs review data packages for suspected faults, high uncertainty, residual abnormalities, phase switching, or periodic sampling windows; the dynamic fusion module combines model confidence, data quality, and residual consistency to generate the final probability; and the alarm state machine performs hysteresis judgment, conflict retention, and timeout degradation.

[0139] The history module saves device identifier, window identifier, operating condition phase, edge model version, backend model version, model probability, dynamic weights, fusion results, conflict or timeout status, root cause hints, and inference time. The front-end interface displays the complete diagnostic chain from edge alert to backend confirmation according to the window identifier.

[0140] Through the above implementation methods, the present invention uses the air compressor operating condition phase, the normal baseline within the phase, and the timing model output together for diagnosis, utilizes the event triggering mechanism to reduce the transmission and calculation of windows that do not require verification, and improves the diagnostic stability under conditions of changing operating conditions, data loss, and communication delay through dynamic fusion and alarm state machine.

[0141] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A rapid fault diagnosis method for subway air compressors, characterized in that, Includes the following steps: S1: Acquire continuous physical quantities, discrete state quantities, and data quality information of the subway air compressor; identify the current operating phase based on the compressor operating status, motor current changes, and reservoir pressure changes; calculate the phase condition residual based on the normal baseline of the corresponding operating phase; and form an enhanced feature sequence containing standardized features, phase condition residual, phase encoding, and data quality mask. S2: Construct a fixed-length sliding window and a phase-aligned window with the working condition phase switching time as the anchor point along the time axis respectively. Input the two types of windows into the first diagnostic model deployed at the edge to obtain the first fault probability, prediction uncertainty and residual abnormality. S3: Determine the background review trigger status based on the first fault probability, prediction uncertainty, residual abnormality and operating condition phase switching event; only when the review trigger condition or periodic sampling condition is met, send the review data packet with device identifier, window identifier, timestamp, enhanced feature sequence and data quality information to the background terminal, and output the second fault probability by the second diagnostic model; S4: Determine the dynamic fusion weights based on the confidence of the two diagnostic models, the data quality of the data received by each model, and the consistency between the model output and residual abnormal evidence, and generate the fusion failure probability; combine the air compressor operating status constraints and alarm hysteresis conditions to output confirmed faults, temporary warnings, or normal results, and retain edge warning records when model results conflict or background verification times out; S5: Associate storage condition phase, window identifier, first failure probability, second failure probability, dynamic fusion weight, fusion failure probability, conflict state, and inference time, and use the window confirmed by the maintenance results as a candidate update sample for data drift detection and model version update.

2. The rapid fault diagnosis method for subway air compressors according to claim 1, characterized in that: In S1, the continuous physical quantities include at least several of the following: compressor pressure, pneumatic panel pressure, cyclone separator outlet pressure, drying tower pressure drop, reservoir pressure, oil temperature, and motor current; the operating phases include at least the shutdown phase, start-up loading phase, steady-state gas supply phase, and unloading phase. Phase switching of operating conditions must be confirmed after the corresponding state conditions have been continuously satisfied by a preset number of sampling points.

3. The rapid fault diagnosis method for subway air compressors according to claim 1, characterized in that: In S1, for each working condition phase and each continuous physical quantity, the median, lower quantile, and upper quantile are calculated based on the samples marked as normal during the training period. The absolute value of the difference between the current standardized feature and the corresponding phase median is used as the numerator, and the difference between the upper quantile and the lower quantile plus a preset positive number is used as the denominator to obtain the phase condition residual. The residual abnormality is generated based on the weighted sum of the residuals of each dimension.

4. The rapid fault diagnosis method for subway air compressors according to claim 1, characterized in that: In S2, the fixed-length sliding window contains 120 to 480 continuous sampling points; the phase alignment window is constructed from the moment when the current working condition phase was last confirmed to switch, and the part that is less than the preset length is padded and a missing mask is set, and the part that exceeds the preset length is resampled or pooled. The two types of windows are spliced ​​along the feature dimension or respectively input into the two input branches of the first diagnostic model.

5. The rapid fault diagnosis method for subway air compressors according to claim 1, characterized in that: In S2, the prediction uncertainty is calculated using the normalized binary entropy corresponding to the first fault probability. When the first fault probability reaches the warning threshold, the prediction uncertainty reaches the uncertainty threshold, the residual uncertainty reaches the residual threshold, or a phase switch of the operating condition is detected, a background review is triggered, and a window is sampled from the window that has not been triggered for background inspection according to a preset period.

6. The rapid fault diagnosis method for subway air compressors according to claim 1, characterized in that: In S3, the second diagnostic model divides the enhanced feature sequence into multiple temporal blocks with overlapping regions. Each temporal block is input into a multi-layer Transformer encoder after linear projection and position encoding. After pooling and fully connected processing, the second fault probability is output. The backend aligns the second failure probability with the corresponding edge end result based on the window identifier and timestamp.

7. The rapid fault diagnosis method for subway air compressors according to claim 1, characterized in that: In S4, the dynamic fusion weight of the first diagnostic model is obtained by multiplying the model's basic weight, model confidence, data quality coefficient, and consistency coefficient between the model output and residual abnormality. The base weight of the second diagnostic model is higher than that of the first diagnostic model, and the dynamic fusion weight of the second diagnostic model is reduced when there are missing data packets, excessively high interpolation ratios, or time alignment errors.

8. The rapid fault diagnosis method for subway air compressors according to claim 1, characterized in that: In S4, the fault triggering threshold is higher than the fault clearing threshold. The system confirms a fault after the probability of a fault reaches the fault trigger threshold a preset number of times consecutively, and resolves a fault after it falls below the fault resolution threshold a preset number of times consecutively. When the background review does not return within the maximum waiting time, a temporary warning is output. After the background result arrives, the confirmation status is updated according to the window identifier, but the original warning record is not deleted.

9. The rapid fault diagnosis method for subway air compressors according to claim 1, characterized in that: In step S5, data drift detection is performed on the feature distribution difference between the recently confirmed normal window and the training baseline; when the drift reaches a threshold or a preset number of maintenance confirmation samples are obtained, the second diagnostic model is updated, and the updated second diagnostic model is used as the teacher model to perform knowledge distillation on the first diagnostic model; version replacement is performed only when the candidate first diagnostic model simultaneously meets the constraints of fault recall rate, model capacity and inference latency.

10. A rapid fault diagnosis system for subway air compressors, characterized in that, include: The system comprises a data access module, a working condition phase recognition module, a phase residual construction module, a window construction module, an edge diagnosis module, a review trigger module, a background review module, a dynamic fusion module, an alarm state machine module, a model management module, and a history record module; these modules are configured to execute the steps of the rapid fault diagnosis method for subway air compressors as described in any one of claims 1 to 9. The edge diagnostic module and the review trigger module are set on the edge computing device at the vehicle end, vehicle gateway or maintenance site, while the background review module and the model management module are set on the operation and maintenance server or cloud computing node. The edge end only uploads review data packets that meet the review trigger conditions or periodic sampling conditions, and sets a non-repeatable window identifier for each review data packet.

Images